Datasetpublic

JamalLee/Omni-Fake-SET

Omni-Fake-SET Omni-Fake-SET is the in-distribution split of Omni-Fake, a unified multimodal deepfake dataset for social-media forensics. It covers image, audio, video, and audio–video talking-head (AV-TH) modalities. Each modality uses the same three-way label space: real, fully synthetic, and tampered. Pair with the held-out benchmark Omni-Fake-OOD for out-of-distribution evaluation. Paper: arXiv:2605.01638 Project page: Omni-Fake License: CC-BY-4.0 Video (hybrid… See the full description on the dataset page: https://huggingface.co/datasets/JamalLee/Omni-Fake-SET.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

5likes2.8kdownloads

Dataset Card

Omni-Fake-SET

Omni-Fake-SET is the in-distribution split of Omni-Fake, a unified multimodal deepfake dataset for social-media forensics. It covers image, audio, video, and audio–video talking-head (AV-TH) modalities. Each modality uses the same three-way label space: real, fully synthetic, and tampered. Pair with the held-out benchmark Omni-Fake-OOD for out-of-distribution evaluation.

—Paper: arXiv:2605.01638
—Project page: Omni-Fake
—License: CC-BY-4.0

Video (hybrid release)

SET Video totals 260,000 clips: 100,000 real + 100,000 full_synthetic + 60,000 tampered.

Component	Count	Format on HF	Source
Real + Full_Synthetic	200,000	`data/Video/video-set.7z.*` (27 parts, ~114 GB)	GenBuster-200K train split
Tampered	60,000	`data/Video/train-*.parquet`	Omni-Fake PartialEdit

Real / fullsynthetic paths after extracting the `.7z` shards: `train/real/` → `real`; `train/fake/<generator>/` → `fullsynthetic (easyanimate, cogvideox, ltxvideo, hunyuanvideo). Tampered parquet rows use label=tampered, generator=PartialEdit`.

Attribution: Real and full_synthetic video clips are from GenBuster-200K (MIT). Cite BusterX (arXiv:2505.12620) when using those clips.

Quick start

python

from datasets import load_dataset

# Image (default config)
ds_img = load_dataset("JamalLee/Omni-Fake-SET", "image")
print(ds_img)

# Audio
ds_aud = load_dataset("JamalLee/Omni-Fake-SET", "audio")

# Video tampered (parquet; Dataset Viewer)
ds_vid = load_dataset("JamalLee/Omni-Fake-SET", "video")
# Real + full_synthetic: download/extract data/Video/video-set.7z.*

# Audio–video talking head
ds_avth = load_dataset("JamalLee/Omni-Fake-SET", "avth")

Fields (by modality)

Image: image, mask, label, generator, filename, split

Audio: audio, label, generator, filename, split, spoof_intervals

Video (parquet, tampered only): video, label, generator, filename, split

Video (`.7z`, real + full_synthetic): extract video-set.7z.*; map train/real/ → real, train/fake/<generator>/ → full_synthetic

AV-TH: video, label, generator, filename, split

Labels are real, full_synthetic, or tampered (same three-class scheme for Image and Video).

Dataset Viewer

Open the Dataset Viewer tab and keep subset `image` selected (default). Subset `video` previews tampered parquet. Real and full_synthetic video require downloading the .7z archive.

Citation

bibtex

@article{li2026omnifake,
  title={Omni-Fake: Benchmarking Unified Multimodal Social Media Deepfake Detection},
  author={Li, Tianxiao and Huang, Zhenglin and Wen, Haiquan and others},
  journal={arXiv preprint arXiv:2605.01638},
  year={2026}
}